Ingestion Layer
Chain indexers, market/liquidity feeds, contract metadata, and governance events into raw storage.
Execution Architecture
Backend MVP plan: data in, scores out.
This is the minimum architecture needed to deliver reliable, versioned risk outputs.
Nightly Score Refresh
Top 250 Tokens
< 300ms p95 API
Versioned Model Bundle
System Overview
Feature Pipeline
Deterministic feature jobs for wallet graph, liquidity behavior, semantic risk flags, and anomaly vectors.
Model Service
Batch scoring with calibrated outputs and attribution metadata written to a serving datastore.
Scoring API
Read-only endpoints for project score, confidence, decomposition, timeline, and model version fields.
Ops and Monitoring
Job orchestration, data quality checks, drift monitoring, and incident alerting for pipeline failures.
Data Flow
1. Collect
Acquire on-chain transactions, token events, LP states, and contract metadata snapshots.
Acquire on-chain transactions, token events, LP states, and contract metadata snapshots.
2. Normalize
Resolve identities, deduplicate events, and align timestamps into canonical schema.
Resolve identities, deduplicate events, and align timestamps into canonical schema.
3. Feature Build
Generate module-specific features and persist versioned feature sets.
Generate module-specific features and persist versioned feature sets.
4. Score
Run model bundle, compute risk outputs, and attach top-signal attributions.
Run model bundle, compute risk outputs, and attach top-signal attributions.
5. Serve
Publish latest snapshot to API and dashboard with model/version metadata.
Publish latest snapshot to API and dashboard with model/version metadata.
MVP Service Contracts
| Endpoint | Purpose | Key Fields |
|---|---|---|
| GET /v1/projects/{id}/score | Primary score view | score, confidence, risk_band, model_version, scored_at |
| GET /v1/projects/{id}/decomposition | Module explainability | module_scores, weights, contributions |
| GET /v1/projects/{id}/timeline | Recent risk-relevant events | events, timestamps, severity, source_type |
| GET /v1/models/latest | Transparency metadata | version, train_window, benchmark_metrics, changelog |
Execution Phases
Phase A (Weeks 1-2)
Ingestion skeleton, canonical schema, and one-chain historical backfill.
Ingestion skeleton, canonical schema, and one-chain historical backfill.
Phase B (Weeks 3-4)
Feature jobs for wallet graph + liquidity + contract semantics, with reproducible run configs.
Feature jobs for wallet graph + liquidity + contract semantics, with reproducible run configs.
Phase C (Weeks 5-6)
Baseline model scoring pipeline and decomposition artifact generation.
Baseline model scoring pipeline and decomposition artifact generation.
Phase D (Weeks 7-8)
Read-only API and dashboard integration with nightly scheduled refresh.
Read-only API and dashboard integration with nightly scheduled refresh.
Operational Controls
Data Quality Gates
Reject scoring runs when source completeness, lag, or schema validity thresholds are violated.
Model Version Discipline
Every release must include benchmark deltas, feature-set hash, and rollback pointer.
Failure Recovery
Serve last-known-valid snapshot if nightly scoring fails, with status flags in API responses.
MVP focus: reliability and transparent versioning first, model sophistication second.